Search

SPARQL query services that balance processing between clients and servers become more and more essential to handle the increasing load for open and decentralized knowledge graphs on the Web. To this end, Linked Data Fragments (LDF) have introduced a foundational framework that has sparked research exploring a spectrum of potential Web querying interfaces in between server-side query processing via SPARQL endpoints and client-side query processing of data dumps. Current proposals in between typically suffer from imbalanced load on either the client or the server. In this paper, to the best of our knowledge, we present the first work that combines both client-side and server-side query optimization techniques in a truly dynamic fashion: we introduce WiseKG, a system that employs a cost model that dynamically delegates the load between servers and clients by combining client-side processing of shipped partitions with efficient server-side processing of star-shaped sub-queries, based on current server workload and client capabilities. Our experiments show that WiseKG significantly outperforms state-of-the-art solutions in terms of average total query execution time per client, while at the same time decreasing network traffic and increasing server-side availability.

The adoption of Semantic Web technologies, and in particular the Open Data initiative, has contributed to the steady growth of the number of datasets and triples accessible on the Web. Most commonly, queries over RDF data are evaluated over SPARQL endpoints. Recently, however, alternatives such as TPF have been proposed with the goal of shifting query processing load from the server running the SPARQL endpoint towards the client that issued the query. Although these interfaces have been evaluated against standard benchmarks and testbeds that showed their benefits over previous work in general, an evaluation of the effects of the query shapes on query performance of the different available interfaces has never been done. In this paper, we present the results of our in-depth evaluation of query shapes impact on the performance of existing LDF interfaces. Using representative and diverse query loads that are designed to include relevant query shapes and are based on the query log of a public SPARQL endpoint, we stress test the different interfaces and identify their strengths and weaknesses.

Spatial keyword querying has attracted considerable research efforts in the past few years. A prototypical query takes a location and keywords as arguments and returns the k objects that score the highest according to a ranking function. While different scoring functions have been used, how to compare different ranking functions for spatial keyword querying still remains an open question with little investigation. We propose a crowdsourcing-based approach to evaluate and compare ranking functions for spatial keyword search. Given two ranking functions f1 and f2, we use a matrix to model all possible binary questions regarding the different results produced by f1 and f2. We propose a multi-step process to reduce the number of binary questions, identifying the most important questions to ask. Further, we design a crowdsourcing model that obtains the answers to those important binary questions from crowd workers. We also devise a global evaluation process that is able to quantitatively compare f1 and f2 based on a multitude of answers received. According to the results of empirical studies using real data, the proposed approach is efficient and able to draw reliable conclusions in comparing ranking functions for spatial keyword search.

Identifying visited points of interest (PoIs) from vehicle trajectories remains an open problem that is difficult due to vehicles parking often at some distance from the visited PoI and due to some regions having a high PoI density. We propose a visited PoI extraction (VPE) method that identifies visited PoIs using a Bayesian network. The method considers stay duration, weekday, arrival time, and PoI category to compute the probability that a PoI is visited. We also provide a method to generate labeled data from unlabeled GPS trajectories. An experimental evaluation shows that VPE achieves a precision@3 value of 0.8, indicating that VPE is able to model the relationship between the temporal features of a stop and the category of the visited PoI.

In recent years, using cell phone log data to model human mobility patterns became an active research area. This problem is a challenging data mining problem due to huge size and non-uniformity of the log data, which introduces several granularity levels for the specification of temporal and spatial dimensions. This paper focuses on the prediction of the location of the next activity of the mobile phone users. There are several versions of this problem. In this work, we have concentrated on the following three problems: predicting the location and the time of the next user activity, predicting the location of the next activity of the user when the location of the user changes, and predicting both the location and the time of the activity of the user when the user’s location changes. We have developed sequential pattern mining-based techniques for these three problems and validated the success of these methods with real data obtained from one of the largest mobile phone operators in Turkey. Our results are very encouraging, since we were able to obtain quite high accuracy results under small prediction sets.

We demonstrate Crowd Rank Eval, a novel framework for the evaluation of ranking functions for top-k spatial keyword queries. The framework enables researchers to study hypotheses regarding ranking functions. Crowd Rank Eval uses crowd sourcing for synthesizing results to top-k queries and is able to visualize the results and to compare them to the results obtained from ranking functions, thus offering insight into the ranking functions.